I wanted a way to tell whether AI was making work better. I ended up arguing with ticket counts, cost accounting and my own assumptions. This is the method I would test, and the long list of reasons to be careful with it.